A Retrieval-Augmented Generation (RAG) system that enables natural language interaction with scientific PDFs, providing contextual, memory-aware responses and seamless document exploration.

Highlights

• Built a LangChain-based RAG pipeline for contextual, memory-aware LLM responses.
• Integrated FAISS for fast semantic search over Gemini embeddings.
• Designed retrieval, document, and memory chains for history-aware conversations.
• Developed an interactive Streamlit UI for seamless document exploration.
• Added comprehensive tests for all modules.

Tech Stack: LangChain, Gemini, Fastapi, FAISS, Streamlit
Demo: 🔗 Live App on Render

View on GitHub →

A Multimodal classification system designed to detect malicious Bengali, Bengali-English code-mixed, and code-switched memes by leveraging image-text fusion and advanced NLP techniques.

Highlights:

• Conceptualized the project idea and led a three-member research team.
• Conducted an extensive literature review to identify research gaps and define the project scope.
• Curated and preprocessed a labeled meme dataset for multimodal analysis.
• Fine-tuned unimodal and multimodal transformer models, integrating image embeddings for improved text classification.
• Designed and implemented a custom multimodal classification model optimized for the dataset.
• Performed explainability analyses to interpret and validate model predictions.
• Authored a draft paper for submission to a reputed conference.

Tech Stack: PyTorch, Hugging Face Transformers, Pandas, Label Studio, Gemini, LaTeX, Notion, Draw.io, Scikit-learn, Seaborn, Matplotlib

View on GitHub →

Developed a predictive modeling system for early diagnosis and management in the medical field, focusing on Bengali clinical data to improve accessibility for local practitioners.

Highlights:

• Fine-tuned transformer models on a custom Bengali clinical dataset.
• Led a team of four, providing guidance and coding support.
• Developed reusable scripts for data processing and model training.
• Analyzed model performance to identify reasons for failure.
• Authored the initial draft for journal submission.

Tech Stack:
PyTorch, Hugging Face, LaTeX, Notion, Mandaley, Scikit-learn, Seaborn, Matplotlib, Pandas

View on GitHub →